Ganoderma lucidum is widely used in the medicine and healthcare industries as a valuable medicinal herb, which is usually artificially grown. Its quality is influenced by various factors, and traditional evaluation methods for this herb are complex. Herein, a combination of short-wave infrared hyperspectral imaging and machine learning methods was explored to develop a prediction model for different ganoderic acids contained in Ganoderma lucidum. First, we established Back Propagation Neural Network (BPNN), Extreme Learning Machine (ELM) and Decision Tree (DT) models using the original spectral data as baseline models; ELM showed the best prediction. Next, we preprocessed spectral data using different preprocessing methods and selected characteristic wavelengths via Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA) and Uninformative Variables Elimination (UVE). The optimal results demonstrated that the MSC–CARS–ELM model achieved the best performance for predicting ganoderic acids A, B, and C2, with prediction correlation coefficients (R²P) of 0.91, 0.88, and 0.91, respectively. This finding confirms that preprocessing combined with feature wavelength selection significantly enhances prediction accuracy. Further analysis revealed that while the BPNN and ELM models generally benefited from these techniques, the DT model showed limited improvement. Hence, short-wave infrared hyperspectral imaging can be used for the accurate, fast, and non-destructive detection of ganoderic acids in Ganoderma lucidum. However, this study still exhibits several limitations. Consequently, future research should expand the sample size, incorporate external validation sets, and explore additional model parameter configurations. Keywards： ganoderic acids, machine learning, hyperspectral imaging, wavelength selection, preprocessing methods • Integrating CARS wavelength selection with ELM creates a highly accurate and efficient model for quantifying ganoderic acids. • Advanced preprocessing and feature selection dramatically streamline the spectral data, focusing on the most relevant chemical information. • The research validates a practical, non-destructive approach that aligns with the needs of industrial quality control for traditional medicines. • The developed methodology establishes a transferable blueprint for ensuring the quality and authenticity of a wide range of medicinal plants.
Ran et al. (Sun,) studied this question.